Every guide on this query is an opinion post. This one is a count: what two assistants did with the same twelve questions on the same day.
How to rank in ChatGPT: there is no rank, and that changes the job
ChatGPT does not publish a ranking, so there is no
position to climb — each answer is assembled from pages it retrieves at
that moment, and your job is to be one of the pages it can retrieve and wants to
quote. In the run behind this page, ChatGPT answered 12
questions about one local market and attached 44 links from 25
different domains — 3.7 per answer, chosen fresh each time. Claude
answered the same 12 questions the same day and attached
0 links after reading 277 pages. Four
things move the outcome, in this order: whether the search crawler can fetch
your page at all, whether the sources the assistant already reads mention you,
whether your own page states its facts in plain server-rendered text, and
whether the assistant can tell your business apart from a similarly-named one.
Everything below is what we measured, on a date, with the transcripts kept.
What one measured run showed
ChatGPT
Claude
Questions asked
12
12
Answers with at least one link
12
0
Links shown, total
44
0
Distinct domains linked
25
0
Pages read to write the answers
820
277
Model
gpt-5-6-sol
claude-sonnet-5
The two engines do not behave the same way, so advice written for one
does not transfer to the other unedited. Claude read more pages than
ChatGPT cited links and surfaced none of them. Any tactic whose whole mechanism
is “get cited” is, on this evidence, a ChatGPT tactic. What both have
in common is earlier: they have to be able to fetch your page.
One run, one market, one day, each question asked once. It is
enough to show the shape and not enough to call any of it a rate — which
is exactly why the paid version asks each question three times.
The four things you can actually change
Let the right crawler in. An assistant’s search
crawler and its training crawler are different user agents with different
consequences, and blocking the wrong one removes you from the answers while
leaving the training set exactly as it was. The
crawler user-agent list names all eleven, with
each vendor’s own documentation linked. Start here, because until this is
right nothing else you do reaches the engine.
Get onto the pages it already reads. In this run the links
went overwhelmingly to businesses’ own websites, with a short list of
review platforms and one forum thread making up the rest — the full
breakdown is in what ChatGPT cites.
That list is short enough to work through by hand, and it is different in every
category.
Put the facts in the HTML. Retrieval reads what the server
returns. Prices, services, hours, address and the answers to your common
questions belong in the source, not behind client-side rendering — a
curl of your own page is the whole test.
Be one entity, not two. 4 of the 12
ChatGPT answers in this run were about a similarly-named business in another
country, answered with full confidence. Consistent name, address and phone
across your own site and the major profiles is what stops that.
How to check whether any of it worked
Measure the way a customer searches, not the way you search for yourself:
Ask questions with no business name in them. Discovery
questions — “best dentist in Ballard Seattle” — are the ones where a stranger
finds you or does not. Brand questions flatter you and tell you almost
nothing.
Ask each question at least three times. The same question
comes back differently; one sample tells you whether you appear at all, not how
often.
Keep the answer, not a score. Full text, every link, the
model version and the timestamp. A number you cannot trace to a transcript is a
number you cannot check — ours are traceable, and so should anybody
else’s be.
Watch the AI Overview separately. In a hand-check of
21 commercial keywords in a different category, 17 showed a
Google AI Overview above the organic results; the
trigger table has the method and the
caveats. Ranking first under an AI answer is not the same as being the
answer.
What nobody can tell you
No one controls the position. The engines decide what to
mention. Anyone selling a fixed spot in an AI answer is selling something they
do not control.
Advice does not transfer between engines. On this evidence,
a citation strategy that works on ChatGPT does nothing on Claude, which cited
nothing at all.
Two engines is not five. We have live-tested ChatGPT and
Claude. Gemini, Perplexity and Google AI Overviews are not connected here, and
where a page of ours talks about them — such as
the Perplexity guide — it says which
part is measured and which part is read off the vendor’s documentation.
Questions about ranking in ChatGPT
Can you rank number one in ChatGPT?
There is no numbered ranking to occupy. ChatGPT assembles each answer from pages it retrieves at that moment, so the useful question is whether it can reach your page and whether the sources it already reads mention you.
What is the first thing to fix?
Crawler access. ChatGPT's search index is fed by OAI-SearchBot, which is a different user agent from the GPTBot training crawler. A robots.txt rule meant to opt out of training that disallows both tokens also takes you out of ChatGPT's search results, so read your own file and check which tokens it actually names.
Does being cited by ChatGPT mean the same thing on Claude?
No. In our run Claude attached 0 links across 12 answers while its transcripts show 277 pages read. A tactic aimed at citations is a ChatGPT tactic on this evidence.
How often should I re-check?
Often enough to see movement rather than noise. Answers vary between runs, so a single check is a snapshot; the same question set asked again on a later date is what turns it into a trend.
Do I need a tool for this?
Not to start. Ask your own discovery questions in the assistant, keep the answers, and note who does get named. A tool is worth it when you want the same questions asked repeatedly, archived, and compared over time without doing it by hand.
Where these numbers come from
Our own run of 15 August 2026: 12 ChatGPT answers and 12 Claude answers to the same 12 questions, archived as raw.jsonl with full text, links, model version and timestamp per answer. The counts on this page are computed from those files at build time.
Our AI Overview hand-check of 21 keywords on 15 August 2026, recorded with the exact query string used; the method and its limits are on the trigger page.